Decoding carbon information disclosure with NLP techniques: Combating carbonwashing for energy and climate transition

Amid the pressing need for carbon information disclosure (CID) and the widespread issue of "carbonwashing" in the process of energy and climate transition, our study comprehensively deconstructs CID from the perspectives of completeness, consistency, readability, and tone. Utilizing a sample of 8578 firms that publish social and environmental responsibility reports, our empirical investigation leverages a suite of natural language processing techniques, including the Doc2Vec algorithm and BERT model, alongside econometric methods such as seemingly unrelated regression and Heckman selection. Our analysis reveals several key findings: firms engaging in CID exhibit selective disclosure, weak consistency and readability, and a marked preference for positive language. There exists limited interrelation between completeness and other CID features, which does not significantly strengthen even with improvements in the external information environment. Our paper not only advances the understanding of how to quantify CID performance but also has significant practical implications for promoting CID quality and addressing "carbonwashing" in China and possibly in other emerging countries where third-party carbon information is lacking.